Zero-shot burned area mapping with the Segment Anything Model (SAM): a label-free framework for post-fire environmental assessment.

Accurate and rapid mapping of burned areas is critical for understanding the impacts of forest fires on ecosystems, the carbon cycle, and post-fire recovery processes. Traditional pixel-based approaches often provide limited accuracy due to spectral confusion, topographic variability, and sensitivity to empirically defined threshold values. Although deep learning frameworks such as U-Net, DeepLab, and SegFormer have achieved substantial improvements in spatial accuracy in recent years, their operational scalability remains constrained by the need for labeled training data and region-specific retraining. This study proposes a zero-shot burned area mapping approach based on the Segment Anything Model (SAM) using Sentinel-2 data. SAM is a foundation model trained on more than one billion segmentation masks and is capable of performing prompt-based segmentation without task-specific additional training. In the proposed methodology, composite images derived from Sentinel-2-based NBR, NBR2, and NDVI indices were generated and used as the primary input representation for SAM under a zero-shot configuration. Subsequently, the effects of different pre-processing, post-processing, and hyperparameter configurations on model performance were systematically investigated. Quantitative analyses revealed that multi-scale configurations, particularly crop_n_layers = 2, substantially improved geometric integrity and boundary accuracy. The highest Intersection over Union (IoU) values were 0.89 for the Bursa site and 0.87 for the Çanakkale site, while the corresponding F1 scores were 0.94 and 0.92, respectively. These results indicate that SAM can achieve performance levels comparable to supervised models trained with labeled data, despite requiring no task-specific training. Furthermore, integrating NBR-, NBR2-, and NDVI-based composite inputs with SAM outputs improved the discrimination between burned and unburned surfaces, particularly by reducing boundary fragmentation and spectral confusion in visually heterogeneous areas. The proposed method reduces dependence on manual labeling, which is one of the major limitations of deep learning-based remote sensing workflows, by operating without fully labeled training datasets. Therefore, it offers a fast and scalable solution with reduced processing costs and strong adaptability to different ecosystems and sensor conditions. In this regard, this study represents one of the first systematic applications of SAM for burned forest area detection and provides a label-free, flexible, and generalizable alternative to conventional supervised deep learning approaches. Overall, the findings demonstrate that SAM can serve as a powerful, scalable, and low-cost framework for zero-shot environmental monitoring and automatic burned area detection, particularly in data-scarce or time-critical post-fire assessment scenarios.